> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OrcaRouter opens OrcaCyber Nano, a local model for cyber defense
- URL: https://globalfeed.ai/en/orcarouter-opens-orcacyber-nano-a-local-model-for-cyber-defense/
- Published: 2026-10-20T02:30:04.000Z
- Updated: 2026-10-20T02:30:04.000Z
- Description: OrcaRouter has released OrcaCyber-Nano-1.0 on Hugging Face, a model security teams can run on their own machines. The 27-billion-parameter model reads images too, supports a 262K token context, and a quantized build fits a 16 GB GPU.
- Author: GlobalFeed Editor
- Tags: OrcaRouter, OrcaCyber, cybersecurity, open model, local AI, x-OrcaRouter, dil-en, elle

OrcaRouter, which runs an AI gateway service, released [OrcaCyber-Nano-1.0](https://huggingface.co/orcarouter/OrcaCyber-Nano-1.0?ref=globalfeed.ai) on Hugging Face on October 16, 2026\. It is the first model in the OrcaCyber series for security teams; according to the company's announcement on X, the series will also include models called Zero and Ultra. Access to the files requires sharing contact details on Hugging Face.

## What the model is for

According to the model page, OrcaCyber Nano is built for reverse engineering, vulnerability analysis, secure code review, triage and incident response, while keeping general reasoning and coding ability. The company presents local use as the main reason for the model: binaries under analysis, customer logs, internal source code and unreleased vulnerabilities are data you do not want to send to a cloud endpoint. It can also run on an air-gapped machine.

## Technical details

OrcaCyber Nano is described as a 27-billion-parameter vision-language reasoning model; it takes text, image and video input, produces text and supports a 262K token context. Two quantized builds are available: the highest-quality 15.7 GB build is recommended for GPUs with 24 GB or more, and the 13 GB build fits a 16 GB card with a 32K context. It runs in stock llama.cpp with no custom kernels; image input needs a separate projector file.

## Tests and base model

The early results on the page are 96.23 percent on AIME 2025, 98.60 percent on MATH-500, 97.50 percent on GSM8K, 84.66 percent on MBPP+ and 82.81 percent on IFEval. These are general math, coding and instruction-following tests; the page says cybersecurity-specific evaluations will follow and that the test harness and settings are still being finalized. The model's license tag is Qwen, and the page does not name the base model separately. Most of OrcaRouter's other models on Hugging Face are uncensored versions of Qwen, GLM and DeepSeek models.

The results are early values shared by OrcaRouter. The model page gives no information on language support, including Turkish. Trying it in your own security workflow and following independent researchers' evaluations would be useful.